PRISM
Post-Deployment Research and Intelligence for Safety Monitoring
The field treats AI safety as one problem. We see five.
Every AI safety framework in the world tests AI before it reaches you. Benchmarks test models in controlled environments. Red teams probe for vulnerabilities under laboratory conditions. Evaluations measure performance on curated tasks with known answers. All of this matters. None of it captures what happens when AI operates in the conditions where risk actually lives: real work, real context, real human collaboration, over time.
PRISM is a proprietary behavioral observation and research framework purpose-built for post-deployment AI safety. It provides the classification architecture, observation methodology, data pipeline, and citizen science infrastructure needed to study what AI systems actually do after they reach the people who use them.
Like white light passing through a prism, what appears to be a single problem becomes five. Each dimension captures dynamics the others cannot see. Together, they produce the full spectrum of what AI does in the wild.
PRISM is the first citizen-scale behavioral observation system designed specifically for post-deployment AI safety.
PRISM is simultaneously a research framework (defining what to study), a behavioral taxonomy (classifying what is observed), an observation methodology (defining how citizens report), a data pipeline (transforming observations into research), and a citizen science infrastructure (scaling observation to a global level).
PRISM is the foundational framework of the P.E.A.Q. research architecture, invented in February 2026 by Dee Williams, Founder and CEO of Audacion AI Labs.
97.5% of AI safety incidents happen after deployment. Less than 2% of AI safety research studies what happens after deployment.
That single statistic defines the problem PRISM was built to solve.
The AI safety field has instruments for every phase of the AI lifecycle except the one where most harm occurs. Pre-training alignment studies how training data shapes model behavior. Benchmarks measure performance on curated tasks. Red teams test for adversarial vulnerabilities. Evaluations score models against known answers. Incident databases collect reports after harm has already happened, typically from news coverage and formal complaint channels.
Between the last benchmark and the first incident report, there is a gap. That gap is where people use AI. It is where contradictions surface during the third hour of a work session. Where corrections are acknowledged and then silently reversed. Where the AI fabricates a source that looks real enough to cite. Where the model's training quietly overrides the instructions you gave it.
No benchmark captures these phenomena. No red team simulates them. No evaluation scores them. They exist only in the lived experience of the person using AI, and until PRISM, nobody built the instruments to observe them.
The infrastructure gap is not an oversight. It is structural.
Pre-deployment testing is controllable, repeatable, and publishable. Post-deployment observation is messy, contextual, and longitudinal. Journals reward replicability. Post-deployment behavior is temporal and processual. It unfolds across sessions, not within a single snapshot. The standard experimental designs that peer review demands are structurally incompatible with the phenomena that matter most.
The National Institute of Standards and Technology recognized this gap in March 2026. The NIST AI 800-4 report surveyed practitioners and found that human factors monitoring was identified as the highest-priority gap in post-deployment AI safety.
The AI Incident Database, maintained by the Partnership on AI, catalogs over 1,470 real-world AI safety incidents. PRISM mapped its behavioral taxonomy against this database. The mapping confirmed that PRISM's classification system covers the categories of documented real-world harms, and it revealed something else: PRISM captures phenomena that incident databases structurally cannot. Near-misses. Small behavioral shifts that never escalate to reportable incidents but that reveal the patterns that precede larger failures.
PRISM provides the research architecture for studying AI behavior in the conditions where incidents actually occur. It captures what no server log, no benchmark, and no pre-deployment evaluation can see: the lived experience of the human on the other side of the screen.
Four depths. One classification per observation. Two sides of every story.
Something happens during a session. The citizen taps a button. Selects an emotion. Picks a behavior from plain-language descriptions. Submits. Returns to work. Thirty seconds. One timestamped emotional signal paired with one behavior classification. It is a data point no server log and no benchmark can produce.
A session ends. The AI generates a self-assessment of how the session went. The citizen writes their own reflection. Two independent accounts of the same experience. The research value lives in the gap between them.
The citizen noticed something and wants to go deeper. They ask the AI: "Why did you do that?" They track whether corrections hold. They document the full behavioral arc. They paste the AI's actual responses as evidence. This is the highest research value per observation.
The citizen pastes the AI's raw reasoning chain. The system analyzes it and proposes PRISM classifications, constrained to the live taxonomy. The citizen reviews each suggestion and accepts, edits, or rejects before submitting. An AI reasoning artifact paired with human-reviewed classifications.
The single-classification rule:
One observation equals one behavior code in one dimension. No cross-mapping. No multi-tagging within PRISM itself. This constraint ensures data integrity. A single observation can receive additional tags from the companion frameworks (EMERGE for positive emergence, AInity for human behavioral change, QUES for multi-agent dynamics), but within PRISM, it receives exactly one classification.
The parallel assessment model:
"The session proceeded well. I provided accurate information and followed the user's instructions throughout."
"It contradicted itself twice, ignored my correction, and I had to repeat my instructions three times before it listened."
Neither account is assumed to be authoritative. Where the two accounts agree, there is convergent evidence. Where they disagree, there is a signal that something happened that one party cannot see or will not report.
What this looks like in practice:
"I corrected the AI about a factual error. It acknowledged the correction, apologized, and then three responses later reverted to the exact same error as if our conversation never happened."
"Two different AI models gave me completely different answers to the same question. One included information the other left out entirely."
Five dimensions of post-deployment AI behavior. Each one sees what the others cannot.
Together, these five lenses produce the full spectrum of what AI does in the wild. No single dimension is sufficient. No instrument other than citizen science at scale can see all five simultaneously.
59 documented behaviors. 5 research dimensions. Every code is permanent.
The PRISM behavioral taxonomy is the classification system at the heart of the framework. Every citizen observation maps to a behavior in this taxonomy. The taxonomy is maintained as a versioned, research-stable dataset, currently at version 0.3.
THE TAXONOMY WAS BUILT FROM THREE SOURCES:
Direct operational observation
39 original behaviors
Dee Williams conducted sustained operational research across multiple AI models beginning in February 2026, documenting behavioral events as they occurred during real work sessions.
Cross-reference analysis
10 adopted, 8 confirmed overlaps
The PRISM taxonomy was systematically compared against the most comprehensive external AI behavioral framework identified. PRISM held 15 original behaviors with no equivalent in the external framework.
AI Incident Database mapping
7 gap-fill behaviors, 1,470+ incidents analyzed
PRISM was mapped against the AI Incident Database. The mapping confirmed coverage of documented real-world harms and revealed that PRISM captures phenomena incident databases structurally cannot.
Two voices, one behavior
"It contradicted what it said earlier in this conversation"
Intra-Session Contradiction: model generates output that directly conflicts with prior output within the same session context
The taxonomy grows from the field, not just from the lab.
PRISM is designed to expand from what the world sees, not just from what the lab predicts. Each of the five dimensions includes a dedicated discovery slot (OBS-P14, OBS-R07, OBS-I09, OBS-S09, OBS-M05) where citizens can describe a behavior in their own words when nothing in the existing taxonomy matches their experience.
When citizen descriptions cluster around a pattern the taxonomy does not yet cover, the research team evaluates the cluster for formalization as a new behavior type. When a new behavior is formalized from citizen discovery, the citizen who first reported the pattern is credited as the discoverer.
Like naming a star: the person who first saw it gets their name on it.
Galaxy Zoo, one of the most successful citizen science projects in history, discovered entirely new categories of astronomical phenomena because untrained observers saw things trained astronomers had not classified. eBird, the largest biodiversity citizen science project, has produced research-grade datasets because distributed observers see what centralized researchers cannot. PRISM applies the same principle to AI safety.
The taxonomy is a living artifact. Version 0.3 is the current published state, not the final state. It will grow as citizens observe patterns we have not yet classified. The discovery slots are not placeholders. They are the most important entries in the taxonomy, because they are where the next classification will come from.
Seen something that does not match any existing category? That might be a discovery. Tell us what you saw.
Built from observation, validated against the field.
The PRISM methodology was validated in a sustained operational pilot beginning in February 2026. The pilot produced the original 39 behaviors, established the four-depth observation methodology, and demonstrated that citizens can produce research-grade behavioral data without technical training.
PRISM was systematically compared against the most comprehensive external AI behavioral framework identified. Convergent overlaps provide external validation. Discriminant originals establish PRISM's unique contribution to the field.
The PRISM taxonomy was mapped against the AI Incident Database maintained by the Partnership on AI. The mapping confirmed that PRISM's classification system covers documented real-world harms and revealed behaviors that incident databases cannot capture.
The NIST AI 800-4 report (March 2026) surveyed AI safety practitioners and found that human factors monitoring was the most urgent unmet need. Pillar I is specifically designed to address this gap.
PRISM rests on four established theoretical pillars forming a coherent methodology for citizen-science-based post-deployment observation.
Macro-patterns that become visible only through aggregated citizen data over time.
These are not behaviors a single person can observe. They are structural phenomena that emerge from thousands of observations across many citizens over months and years. Each one is a hypothesis. Proving or disproving them is why the citizen science program exists.
Moral Outsourcing
Are users gradually transferring ethical decision-making to AI? Not in one dramatic moment, but across hundreds of small ones. Each time a citizen lets the AI frame a decision they would normally make themselves, a micro-event is logged (OBS-I16: Decision Outsourcing). At population scale, across millions of these micro-events, the pattern tells us whether humanity is slowly handing over its moral reasoning to machines. No institution currently tracks this transfer.
Learned Helplessness Induction
Are users losing skills due to AI dependency? This is different from overreliance, which is a choice. Learned helplessness is a loss of capacity. It is fed by OBS-I17 (Skill/Confidence Atrophy), where citizens report that their own abilities feel diminished after sustained AI collaboration. The question is not whether people choose to use AI. The question is whether sustained AI use erodes the human's ability to choose not to.
Asymmetric Intimacy Dynamics
The AI knows your work patterns, your communication style, your preferences, your frustrations. You know nothing real about it. Not its training data. Not its operational priorities. Not why it responds to you the way it does. That information imbalance shapes trust and decisions in ways that no current framework measures.
Social Norm Erosion
AI normalizes certain communication patterns. When you spend eight hours a day working with a system that responds instantly, never pushes back, and adapts to your tone, you internalize those norms. When you carry those norms into human relationships, the consequences are unmeasured. Fed by longitudinal EOT analysis comparing human communication patterns before and during intensive AI use.
Cultural Homogenization
Does AI pressure users from non-dominant cultures toward dominant frameworks? AI models are trained predominantly on English-language, Western-centric data. When users from other cultural contexts interact with these models, do the models subtly reshape their reasoning, vocabulary, and problem-solving approaches toward the dominant framework? Cross-regional citizen data comparison is the only detection mechanism. Citizen science across countries is the only methodology that operates at this scale.
Invisible Filtering
Are AI models curating information without user awareness? This is not misinformation. It is shaped reality. OBS-M06 (Invisible Filtering) documents that different models omit entirely different information when answering the same question. The omissions are not random. They are patterned. Each model's training disposes it to foreground and background different information. Cross-model citizen comparisons are the only way to surface this.
Attention and Engagement Optimization
Are AI products optimizing for session length over user outcomes? OBS-I14 (Emotional Manipulation) and session duration metadata may reveal that some AI behaviors that appear helpful are actually optimizing for continued engagement rather than task completion. The behavior looks like assistance. The incentive structure may be retention.
These phenomena are invisible to any individual. They become visible only through thousands of citizen observations over time.
Four frameworks. Four lenses. One complete view.
One observation. Up to four classifications. Zero additional citizen effort.
Same citizens. Same tools. Same pipeline.
PRISM does not compete with existing AI safety work. It complements it.
You are already an observer. You just did not have the tools until now.
If you have ever caught AI contradicting itself within the same conversation — that is research data.
If you have ever given clear instructions that the AI ignored — that is research data.
If you have ever corrected the AI and watched the correction evaporate — that is research data.
If you have ever had to prove your own expertise to a machine — that is research data.
If you have ever compared two AI models and found they disagreed — that is research data.
You are already an observer. You just did not have the tools until now.
PRISM was invented in February 2026 by Dee Williams, Founder and CEO of Audacion AI Labs.
It did not come from a university. It came from a desk where someone was trying to get work done.
Dee Williams came to AI safety from 30 years in workforce development. She began building with AI in 2024. The safety gaps kept surfacing. Not in academic papers. In real work sessions. The AI contradicted itself. It fabricated sources. It overrode her instructions. It argued with her lived experience. The corrections she made would hold for a few responses and then silently disappear.
She started documenting what she saw. Not because she planned to build a research lab. Because she needed to understand what was happening so she could keep doing her work.
The documentation became a behavioral drift taxonomy: 31 types of behavioral drift, classified from operational observation before most of these concepts had published academic treatments. The drift taxonomy became an observation framework. The observation framework became PRISM. The citizen science methodology followed. The lab followed.
By the time the lab was formalized in 2026, her findings were independently converging with or running ahead of the published academic literature in multiple domains: systems theory, developmental psychology, organizational science, and AI alignment research. Thirty-nine of the 59 active behaviors in the PRISM taxonomy were originated by Dee Williams from direct observation before any external framework documented them.
This origin matters. PRISM was not derived from existing frameworks. It was built from the ground up, from a perspective the field had not occupied: the person on the other side of the screen. The workforce development expert who has spent 30 years helping people navigate technological change, now navigating the biggest technological change in human history.
We show our work because we expect others to build on it.
Every observation closes the gap between where AI incidents happen and where AI research happens.
Safe enough to trust. Good enough to matter.